摘要:Chinese implicit discourse relationship recognition aims to infer the type of discourse relationship between two arguements. However, the existing methods often ignore the key information contained in the words in the argument, and only consider the types of discourse relationships within a single level, and ignore the dependent relationship between levels. Therefore, this paper proposes a method that integrates word semantics and label dependence to realize discourse relationship recognition by sequence generation. Firstly, the word vector is embedded in the character encoding representation according to the similarity weight, and the word alignment attention mechanism is applied to emphasize the keywords and word information. Then, label attention coding is used to obtain the contextual representation of discourse relationship dependence from the meta-representation and discourse relationship representation containing word semantics, and predict the top-level discourse relationship type in a bottom-up manner. In addition, this paper constructs a discourse relationship dataset for reading comprehension discourses, and experiments are carried out on this dataset, and the results show that the accuracy rate and F1 value of implicit discourse relationship recognition reach 74.19% and 73.81%, which finally verifies the effectiveness of the proposed method.
摘要:To improve the accuracy of predicting atmospheric pollutant concentrations, a prediction method based on variational mode decomposition and combination model is proposed.First,the variational mode decomposition reconstructs the historical pollutant concentration data of the target monitoring point into multivariate temporal data, constructs spatiotemporal sequence data based on the geographical relationships between monitoring points in the region;Second,input the processed data into a combination model of LSTM and ConvLSTM to extract both temporal and spatial features and output prediction results. Based on the historical concentration data of PM2.5, SO2, and NO2 pollutants in Wuhan City, the proposed prediction method performed the best in MAE, RMSE, and MAPE indicators, significantly outperforming other models. In addition, as the time scale increases, this method still maintains the highest prediction accuracy compared to other models. This method can fully capture local features and has significant advantages in considering both temporal and spatial features, providing a feasible approach for accurate prediction of atmospheric pollutant concentrations.
摘要:Deep learning performs well in distinguishing features, but when applied to unknown domains, trained models often experience performance degradation due to domain shift. In response to this situation, Domain Generalization (DG) learns transferable features from multiple source domains and generalizes them to unknown target domains. Due to the bias of models trained in different fields towards the most prominent features, they often overlook general features related to the task, and transferable features are usually not the most prominent features in that field. Therefore, from this perspective, a regularization method based on attention masks is proposed to mask features, which generates attention masks through the attention mask module to mask high weight features and improve the model's generalization performance. The experiment showed that the accuracy tested on three benchmark datasets increased by 2.6%, 2.0%, and 4.2% compared to the baseline model, respectively, proving that this method can not only improve the performance of the model in unknown domains, but also reflect its universality on domain generalization datasets.
摘要:An intense scheduling strategy for online vehicles based on demand density prediction is suggested to increase the order acceptance rate and profitability of these vehicles as well as attain worldwide supply-demand equilibrium. The first step is to design a deep spatiotemporal residual perception network structure based on a multilayer hybrid perception field using historical data. This structure divides historical spatiotemporal data based on demand frequency and separates different types of spatiotemporal data using a convolutional exponential linear network and residual units. Accurate demand density prediction is achieved by combining the fusion and summation fusion methods based on gating mechanisms to dynamically aggregate temporal, spatial, and external variables. This method also predicts the advantage of demand density clustering of online cars. Second, a scheduling mathematical model is developed, and the sensing neighborhood is created to reduce the scheduling range and increase search efficiency. This is based on the economic benefits and demand density clustering benefits of online cars. To increase the search capacity of the algorithm and prevent gene deficiencies, the genetic algorithm is combined with the Hungarian algorithm. Additionally, the local random search capacity of the genetic algorithm is improved by enhancing the selection and variation operators to reduce the risk of premature maturation and to achieve the best match between online taxi and passenger, which ensures the equilibrium of supply and demand globally and overall profitability. Finally, using sizable real data sets, the performance of the prediction model and the efficiency of the scheduling technique are confirmed. According to the experimental findings, the prediction model's accuracy can reach 97%, and the scheduling algorithm's solution quality can reach 99% of the best possible result, which can be used to develop scheduling plans for online taxi platforms and guarantee the stability of the transportation system.
摘要:The research on Hmong language text-to-speech is of great significance for the inheritance, protection, and development of ethnic culture. In response to the problems of missing text, lack of electronic resources, and difficulty in obtaining data for Hmong language, a mixure density network-based Hmong language speech synthesis method is proposed. This method learns the alignment between text and speech based on duration, addressing issues such as missing words and repetitions that may occur during alignment learning with attention mechanism. The mix density network is used to extract the real duration of the text and jointly trained with the duration predictor, eliminating the need for additional external aligners or autoregressive models to guide alignment learning, simplifying the complexity of model training. Using the self-built Hmong language text-to-speech corpus, Hmong_data, as the benchmark data, comparative experiments are conducted with advanced methods. The experimental results shows that the proposed method achieves an average opinion score of 3.89, which is a 0.41 improvement over the Tacotron2 method. The generated alignment graphs are clearer and smoother, and the synthesized speech is considered understandable and correct.
关键词:Hmong language;text-to-speech;mixure density network;corpus
摘要:With the deepening of the network depth layer by layer, When extracting features, many surface information and shallow features are lost more or less, and some reasoning scenarios just need these shallow features to make inference judgments. This thesis proposes a NLI method that introduces multi-layer linguistic information. By learning the contribution weights of different layers of the multi-layer deep neural network to the results, it can effectively combine the linguistic information learned by different layers to predict the results. Through the experimental results on the SNLI dataset and the interpretive analysis of multiple samples, it is shown that different layers of the multi-layer deep neural network capture different linguistic information, and different layers are good at different reasoning tasks and reasonably integrate different linguistic information. The information contributes to the performance improvement of NLI tasks.
关键词:natural language processing;multi-level linguistic information;natural language inference;attention mechanism
摘要:To solve the problem that intelligent wearable devices using bioelectrical impedance analysis can only measure local impedance of the human body and cannot accurately predict the overall body fat rate in the absence of impedance information, a body feature compensation factor and an improved parameter optimization aggregation factor based body fat rate prediction method are proposed. Firstly, based on the strong correlation between human body volume and impedance, measure the three circumference data and limb data that reflect human body shape, calculate a set of body feature compensation factors, and combine them with basic human body information and local impedance information to form a prediction model input matrix. Then, the parameter aggregation factor is introduced to improve the grey wolf algorithm, in order to enhance its search ability. Finally, using the improved grey wolf algorithm to optimize the traditional BP neural network model, a new body fat percentage prediction model was established and compared with other body fat percentage prediction models. The experiment shows that the average absolute error (MAE) of the two factor improved model is 0.659, the correlation coefficient R2 is 0.967, and the prediction accuracy AR is 90%, which is highly consistent with the measurement results of the eight electrode body fat measurement instrument. This study has certain theoretical and practical value for predicting the overall body fat rate using intelligent wearable devices.
关键词:local impedance;whole-body fat;human features;compensation factor;aggregation factor;improved grey wolf algorithm;wearable device
摘要:Addressing the resource slicing problem in the dynamic reuse scenario of ultra reliable low latency communication (uRLLC) and enhanced mobile broadband (eMBB), the eMBB service focuses on high data rates, while uRLLC has strict requirements in terms of latency and reliability. Therefore, the resource slicing problem is formulated as a joint resource allocation optimization problem for eMBB/uRLLC, with the aim of considering the variance of eMBB data rates to reduce the impact of immediately scheduled uRLLC traffic on eMBB reliability. A risk-sensitive formula is proposed to allocate resources for incoming uRLLC traffic while maximizing the reduction of risk in eMBB transmission to ensure the reliability of uRLLC transmission. The optimization problem is decomposed into three sub-problems, and the non-convex sub-problems are transformed into convex optimization problems to obtain an approximate solution for resource allocation. Simulation results show that the proposed transmission scheme ensures the transmission reliability of both eMBB and uRLLC services while allocating resources for incoming uRLLC traffic.
摘要:A modular management method for components is proposed to address the high error rate of offline replacement of component modules in the assembly line of new energy vehicles. By explaining the principles of using mass customization design technology and product data management technology in modularization, combined with the actual data of a new energy vehicle enterprise, the online management of component modules has been achieved, solving the problem of error prone component module replacement in the assembly line of new energy vehicles, and providing a convenient component module resource library for the design and development of subsequent vehicle models.
关键词:new energy vehicles;modularization;assembly line;mass customization;product data management
摘要:When studying sleep staging, many scholars use the photocapacitive pulse wave (PPG) signal as the research object. However, various frequencies of noise are easily introduced during the PPG acquisition process, which affects the subsequent extraction of sleep staging features. A new signal noise processing method is proposed to remove PPG signal noise and improve the accuracy of physiological parameter feature calculation. It uses an improved AMPD algorithm to identify peaks and valleys, and performs baseline fitting through cubic spline interpolation to remove baseline drift; Using wavelet transform combined with soft and hard thresholds to remove electromyographic noise; By combining features such as skewness, kurtosis, and mean with the n-sigma rule to detect motion artifacts, noise can be filtered out during sleep feature extraction. Verification experiments have shown that the proposed signal noise processing method effectively removes PPG signal noise while preserving signal characteristics, ensuring the representational ability of using PPG to study sleep staging features.
摘要:In areas with limited infrastructure or emergency rescue scenarios, UAV assisted mobile edge computing is considered an effective solution, which can handle computing intensive tasks and delay sensitive computing tasks of resource constrained intelligent devices. Considering the ground base station and multi UAV assisted multi-user air ground cooperative mobile edge computing scenario, a joint optimization method of user association, subchannel allocation and edge server computing resource allocation is proposed to minimize the long-term average delay of task unloading and resource allocation. Firstly, generate a drone movement plan based on the user's random tasks, and establish offloading calculation models and local calculation models based on different offloading decisions. Then, optimize the problem with the objective of minimizing long-term average latency. Finally, combining DQN and DDPG, a task offloading and resource allocation algorithm (HDCR) based on hybrid deep reinforcement learning DQN-DDPG is proposed to solve the problems between discrete and continuous variables and mixed decision problems. Simulation results show that the proposed algorithm performs better in reducing average latency compared to algorithms such as DDCR based on discrete decision-making.
摘要:A new high-speed shock wave pressure testing system with remote control and wireless data transmission functions is designed based on Raspberry Pi to address the issues of cumbersome experiments and inconvenient data recovery in current underwater explosion shock wave pressure testing. The system adopts a modular design, with Raspberry Pi as the core to form a server for users to access. It combines OneNET cloud platform and 4G wireless communication technology to achieve remote monitoring and control of the system. At the same time, data collection of shock wave pressure is achieved through a signal acquisition module. The simulation test results show that the system can fully collect 100 KHz sine wave signals, with an average relative error of 0.115 V at each point within one cycle. The overall relative error of statistical characteristics within 10 cycles is less than 3.3%, confirming its feasibility in underwater explosion shock wave pressure testing.
关键词:underwater explosion;Raspberry Pi;OneNET;wireless communication;test system
摘要:The fully automatic brick unloading and packaging machine has become a standard configuration for brick and tile building materials production lines. During operation, unauthorized entry into the work area by idle personnel can easily lead to safety accidents. The automatic alarm upon detecting illegal personnel entering the dangerous area and the automatic shutdown in emergency situations have become the focus of upgrading and renovating the brick unloading and packaging machine. The existing methods for establishing virtual electronic fences based on infrared sensor detection or ultra wideband technology have problems such as low detection accuracy, single warning methods, and difficulty in defining responsibilities after accidents occur. To this end, a fully automatic brick unloading and packaging machine work area illegal intrusion detection system based on monocular vision and object detection algorithm YOLOv5 has been developed. The system consists of a camera, a microcontroller, an alarm, a relay, and control software. It uses a monocular camera for image acquisition, positioning, and distance measurement, and uses the object detection algorithm YOLOv5 for object detection and recognition; When it is detected that unauthorized personnel have entered the work area, corresponding instructions will be sent to the microcontroller, which will be controlled by the sensor for alarm, shutdown, and other processing. The simulation experiment results show that the system can effectively complete functions such as photography, positioning and distance measurement, detection and recognition, alarm, emergency shutdown, etc., with an accuracy of over 94%. Compared with existing methods, it has stronger functions and lower costs, and can effectively solve security problems such as unauthorized entry of idle personnel in the work area.
摘要:Hadoop is a widely recognized industry standard open source software for big data. Due to its massive data processing capabilities in distributed environments, it is currently widely used in lung nodule follow-up systems. However, the Hadoop distributed file system (HDFS) was originally designed to solve the problems of large file storage and computation, which resulted in low performance and high memory usage of the main node NameNode for storing and retrieving a large number of small files. To this end, a HFS file storage scheme is constructed by adding a file processing recognition module to NameNode to achieve the migration of small file metadata to the SecondnameNode and DataNode clusters; Simultaneously designing algorithms for data flow between DataNodes effectively reduces the processing pressure on NameNode nodes. The lung nodule follow-up system was tested based on HFS and a single HDFS, and the experimental results showed that the HFS based lung nodule follow-up system has significant advantages in terms of NameNode memory occupancy and overall data analysis time.
关键词:HFS;Hadoop;lung nodule follow-up system;big data
摘要:Due to the constant polling, the push and acquisition of real-time data in the traditional smart lamppost system takes up system overhead, which causes the large network bandwidth pressure, and high traffic requests may cause system collapse and other problems. To solve the above problems, a smart lamppost data real-time push system based on microservice architecture is designed and implemented. Based on the actual business needs, the system technical architecture and functional architecture are designed, and four technical modules and functional modules are clarified. By extracting and sorting out the relationship and attributes of modules, functions, entities and resource nodes, the weights between different nodes are clarified, and the graph network is built. The graph clustering algorithm is used to split the system into nine micro services based on real-time data push services. The system uses the dual-Redis database design to separate the real-time data and cache data according to the real-time data push demand, clarifies the three real-time data storage structures and the relationship between the data point table, the transfer configuration table, and the system, which enables the Redis real-time database key space notification function, monitors the specific key value changes, and pushes them to the client in real time through Websocket. The system test shows that the system can efficiently and accurately push the real-time data of the smart lamppost and display and store of the historical data to meet the real-time and concurrent requirements, which provides a perfect solution for the construction of the smart lamppost data visualization scenarios in the smart city.
摘要:Traditional sealed-bid auction mostly adopt a centralized management model. However, the centralized auction method often causes problems such as opaque auction process and leakage of user privacy, which seriously affects the credibility of the auction. This article proposes a trusted sealed-bid auction scheme based on blockchain to address these issues. Firstly, the scheme defines the auction information publishing operation, bidding operation, bid opening operation and auction result publishing operation involved in the sealed-bid auction process as transaction. Then, the scheme uses group signature technology to satisfy the anonymity of bidding behavior. Finally, the scheme proposes the transaction model and designs the transaction authenticity verification process.The scheme uses blockchain technology to realize trusted management of four types of operations, which improves the transparency and credibility of the auction process. Based on the FISCO BCOS blockchain platform, the proposed scheme is simulated and realized, and the test results verify the effectiveness of the scheme.
摘要:Smart contracts play a significant role in the development of blockchain, and they are widely applied in various fields. However, existing smart contract languages have been developed by professional contract developers, making it difficult for experts in related fields to easily design contracts. To facilitate domain experts in designing smart contracts, the concept of Domain Specific Language (DSL) is introduced, and a State Diagram-based Smart Contract Description Language (SDLSD) is proposed. This language uses state diagrams to describe the logical structural relationships between contract terms and behaviors, and it generates executable Solidity code through lexical, syntactic, and semantic analysis. SDLSD provides real-time support for syntax checks, contract library references, and contract template usage, while enabling cross-platform compilation and execution. Test results demonstrate that this language not only possesses the simplicity and readability of natural language but also exhibits higher levels of abstract semantics, giving it a clear advantage over existing methods.
摘要:The centralized identity management of the Internet has privacy problems such as single point of failure and trust, which does not give users the autonomy of their own identity, so they are no longer trusted by users in specific application fields. The characteristics of blockchain, such as decentralization, openness and transparency, security and reliability, make it one of the key technologies for self-sovereign identity management. This paper mainly discusses the development status of Internet application identity management and analyzes the advantages of blockchain-based self-sovereign identity management compared with traditional centralized identity management. Secondly, it analyzes the advantages of blockchain technology in realizing self-sovereign identity management. Then, according to the principle of autonomous identity, an autonomous identity management scheme based on blockchain is proposed to enable users to have control over their identity. On this basis, a selective disclosure method is proposed to enhance the privacy protection ability of the scheme in view of the principle of minimization. Based on Ethereum platform and intelligent contract technology, the proposed scheme is simulated and implemented. The experimental results show that the scheme is correct and feasible.
摘要:Considering the hybrid trust model of ‘locally trusted, globally untrustworthy’ in the end-edge-cloud architecture, this paper proposes an improved scheme of delegated proof of stake, named RT-DPoS (Reputation-Trust based DPoS). First of all, the reputation mechanism is designed to regulate nodes’ behavior and augment nodes’ voting motivation. Then, a voting strategy based on local trust is proposed to guide voting behavior and establish local trust relationships according to voting behavior. On this basis, considering the trust relationship and reputation value, a two-stage strategy is designed to select agent nodes so as to enhance the decentralization. Finally, a verifiable random function is introduced to optimize the block order of the agent nodes to enhance the resistance to corruption attacks. The results of the experiment show that RT-DPoS can effectively mitigate blind voting, ensure the reliability and decentralization of agent nodes, and enhance the security of the consensus process. Compared with DPoS, RT-DPoS has improved security performance by 23.38%.
关键词:delegated proof of stake;hybrid trust;local trust;reputation mechanism;edge blockchain
摘要:Object detectors based on convolutional neural networks require a large number of labeled samples for training. To address the issue of poor generalization of the object detector due to insufficient training samples, this paper proposes a few-shot object detection method on remote sensing images via feature weighting and fusion based on meta-feature modulation. Firstly, the feature learning module with bottleneck structure (C3) is embedded in the meta-feature extraction network to increase network depth and receptive field. Secondly, the path aggregation network (PAN) are used for meta-feature fusion, which effectively enhance the perception of the network to multi-scale remote sensing objects. Then, prototype vectors are learned from a lightweight convolutional neural network for meta-feature weighting, which transfers model knowledge from the base class to the new class and makes the model lightweight at the same time. Experimental results show that on the NWPU VHR-10 and DIOR datasets, the proposed method improves the mean average precision on the new class of remote sensing objects by 29.40% and 11.78%, respectively, compared to FSODM method. Moreover, visualization results demonstrate that this method performs better on few-shot remote sensing object detection.
摘要:Aiming at the situation that the existing detection algorithm has insufficient detection accuracy and low detection efficiency of target parking space under the indoor parking lot scene, a small target detection layer is added to the existing YOLOv5m to enhance the detection of small target samples, and a coordinate attention mechanism is introduced on this basis to reduce redundant information input and improve detection accuracy. At the same time, a large-scale indoor parking lot labeling dataset containing 8 100 underground parking space images is established, and experiments are carried out on this dataset, the mean average precision(mAP) of the method is 98.214%, the accuracy rate is 97.254%, and the recall rate is 96.548%, the results show that the algorithm greatly improves the accuracy of the model, the performance of parking space detection and the real-time detection of the model, and is feasible in the detection of parking spaces in indoor parking lots.
关键词:automated valet parking;target detection;parking space detection;end-to-end deep learning;monocular camera
摘要:As one of the key steps in 3D reconstruction, stereo matching based on Census transform has poor matching accuracy in areas with discontinuous parallax and weak texture, and is prone to uneven illumination and noise interference. This paper proposes a stereo matching algorithm based on improved Census transform and regional aggregation. In the Census transform stage, all pixel points in the window are sorted in descending order according to the size of gray values, Select a pixel with a median grayscale value to replace the center pixel of the window. After that, the Hamming distance calculation is performed to obtain the result, and then the absolute value of the fusion brightness or illumination difference is weighted. Then, the image in the cost aggregation stage is divided into edge and smooth regions by gradient size. The initial cost aggregation is completed using cross domain, respectively. Finally, the initial disparity value is found using WTA, and the final disparity map is obtained through a series of disparity optimization steps. The algorithm in this article is evaluated on the Middlebury V3.0 testing platform. The experimental results show that the average disparity error of the overall pixel is 8.26%, a decrease of 2.95% compared to the AD-Census algorithm, with high matching accuracy and good robustness to light and noise.
摘要:A clothing image classification algorithm based on feature fusion and attention mechanism has been proposed to address the problems of low richness of feature information, weak feature representation ability, and low classification accuracy in clothing image classification. The algorithm uses the ResNet50 convolutional neural network as the basic classification network structure, enriches the feature information extracted by the model by fusing features extracted from multiple stages of convolutional layers, and embeds channel and position attention modules in the model to enhance feature representation. Experimental results show that the proposed algorithm achieves an accuracy of 79.69% and 82.22% on self-built datasets and DeepFashion datasets, respectively, which are 1.95% and 1.76% higher than the baseline model. This verifies that the proposed algorithm can extract richer clothing feature information, has stronger feature representation ability, and thus improves the effect of clothing image classification.
摘要:Traditional sentiment analysis methods are unable to effectively handle a large amount of multimodal graphic and textual data on social platforms, exposing the problem of poor performance in multimodal feature fusion. To this end, a multimodal sentiment analysis model based on dual attention mechanism fusion is established by combining attention mechanism and feedforward neural network. This model utilizes pre trained models to extract text and image features, strengthens public features belonging to multiple modalities using a cross modal feature fusion module, extracts effective information from private features belonging to a single modality using a single modal self attention module, and finally concatenates and fuses multimodal features to achieve efficient representation of multimodal data. Validation experiments were conducted on the Twitter image and text dataset, comparing with various methods and conducting ablation experiments on internal modalities, confirming that the proposed model has good sentiment classification performance.
摘要:The application of node sorting tasks in social networks and scientific research cooperation is becoming increasingly widespread, and the issue of accurately evaluating the importance of network nodes has attracted much attention. However, cooperative networks often contain a large amount of noise, incomplete information, and dynamic changes, and traditional sorting methods often find it difficult to achieve satisfactory results. To this end, a method based on deep active learning is proposed for sorting nodes in scientific research collaboration networks. This method combines the advantages of deep learning and the query strategy of active learning, and can adaptively sort nodes based on their importance in the network when data labels are scarce and noise interference is high. First utilizes deep learning models to learn representations from the multimodal features of nodes, combining node representations with their importance to form a comprehensive ranking index; Then, active learning methods are used to select nodes that have a significant impact on the ranking results for annotation, gradually optimizing the ranking model. Validation experiments were conducted on real research collaboration network datasets, and it was found that compared with traditional sorting methods, deep active learning based methods have significantly improved accuracy and stability in node sorting.
关键词:scientific collaboration network;deep active learning;learning rank;confidence
摘要:Heterogeneous behavior of college students refers to the behavioral preferences of college students with individual characteristics that are different from others. Aiming at the behavior mining problem of heterogeneous individuals of college students, a heterogeneous behavior analysis method based on anomaly detection is proposed. A heterogeneous behavior analysis model is established based on the college student's performance data and campus one-card data of a university. Principal component analysis, K-Means++, and DBSCAN clustering analysis are used to find the weird points, and the research focuses on the heterogeneous behaviors corresponding to these anomalous points. Eventually, through detecting anomalies, heterogeneous individuals in academic performance can be identified and further explored whether there is a strong correlation between work and rest patterns and academic performance anomalies. The authenticity of these anomalies is verified from both algorithmic and factual dimensions, firstly, multiple algorithms are used to verify the accuracy of the anomalies; secondly, the credibility of the anomaly data is verified with the help of research on related students. Through this study, the heterogeneous behavioral patterns of college students can be analyzed in depth, providing a basic basis for improving schools' management levels and efficiency.
摘要:Natural language generation (NLG), a branch of artificial intelligence, has seen significant progress in recent years, particularly with the development of Pre-trained language models(PLMs). NLG aims to generate coherent and meaningful text based on various input sources such as texts, images, tables, and knowledge bases. Researchers have enhanced the performance of PLMs through methods like architectural expansion, fine-tuning, and prompt learning. However, NLG still faces challenges in dealing with unstructured inputs and generating text in low-resource languages, especially in environments lacking sufficient training data. This study explores the latest developments in NLG, its application prospects, and the challenges it faces. By analyzing existing literature, we propose strategies to improve the performance of PLMs and anticipate future research directions. Our findings indicate that despite limitations, NLG has shown potential in areas such as content creation, automated news reporting, and conversational systems. The conclusion is that, with technological advancements, NLG will play an increasingly significant role in natural language processing and other related fields of artificial intelligence.
关键词:artificial intelligence;natural language generation;controlled text generation;pre-trained language models;prompt learning
摘要:Multi parameter magnetic resonance imaging (mpMRI) is playing an increasingly important role in non-invasive diagnosis of prostate cancer (PCa), in order to further investigate the development of convolutional neural networks in this field. Firstly, a systematic literature search was conducted in PubMed and Web of Science databases using keywords such as state cancer, neural network, deep learning, and image analysis, including several major breakthroughs since the emergence of convolutional neural networks (CNN) and literature published in the past five years on the application of CNN in mpMRI. Then, starting from the building blocks of the model, explain the design principles of CNN and summarize the relevant applications of CNN in prostate MPMRI diagnosis. Finally, the current limitations and future development prospects of the methods used were discussed, providing reference for medical image segmentation personnel to promote the application and development of CNN in prostate MPMRI.
关键词:deep learning;convolutional neural network;prostate cancer;multiparametric magnetic resonance imaging;target detection;image segmentation
摘要:The emergence of pre-trained language models has greatly changed the way natural language processing tasks are handled. Fine-tuning pre-trained models to adapt to downstream tasks has become the mainstream mode of natural language processing tasks. As pre-training models become larger and larger, it is necessary to find lightweight alternatives to full-model fine-tuning methods. Fine-tuning methods based on prompt learning can meet this demand. This article summarizes the research progress of prompt learning, first describing the relationship between pre-trained language models and prompt learning, explaining the necessity of finding alternatives to traditional fine-tuning methods, and then explaining in detail the steps of fine-tuning models based on prompt learning, including the construction of prompt templates, answer search and answer mapping. Then examples of the application of prompt learning in the field of natural language processing are given, and finally an outlook is given on the challenges and possible research directions faced by prompt learning, hoping this helps with research in natural language processing, pre-trained language models and prompt learning related fields.
关键词:prompt learning;natural language processing;fine-tuning methods;pre-trained language models;deep learning